Personally, my guess is that it’s actually just plagiarizing the training set in a way that most researchers will come to view as a kind of cheating. What I mean by that is, if you take some plagiarism detection software and run it on GPT-3’s output, it will ring like crazy.
I say this both because I believe it and because if it’s not the case, if we really have a proto-AGI on our hands, then being wrong won’t matter. I sincerely hope that we are a thousand years away from that, because otherwise we are plainly doomed.